John Searle's Chinese Room, ChatGPT and TAD

Searle's Chinese Room thought experiment is famous for clarifying the difference between syntax and semantics. Just because a system got the syntax right, does it mean that it truly understood it in the deepest, finest, human way of thinking? In this thought experiment, there is a person locked in a room with absolutely no interest in furthering his knowledge or understanding. Through some window into the room, he receives "Chinese" symbols (and he has no clue what they are as he has no knowledge of Chinese) So he mechanically goes to a table present in the room which has instructions on what to do with these Chinese symbols he received in the room and also, how to further generate even more symbols. The generated symbols are then spit out from the room. He is just an apathetic agency that does this brute-force work.

Eventually, Searle convincingly postulated that one could form "questions" and feed it into the room and out, comes "answers" from the room -- There is no real understanding here. Remember, the person in the room is apathetic. He is merely following instructions kept on the table -- those instructions were previously left by a truly intelligent human.

That thought experiment was quite prescient of John Searle, I would think he captured quite an important part of ChatGPT. There is one difference though, in case of ChatGPT it has a self-correction system. Let me explain that:

In ChatGPT you could ask for a program written in Javascript such as "Give me a program in Javascript that will use the Threejs library to depict 3D views of objects, along with shadows" -- and presto! it will write out the code for you. (Try it. It actually does a neat job!) You get the feeling that chatGPT truly understands.

It does not.

It simply used a set of instructions kept on a table in its room, on how to identify key "syntactic" form of your sentence, break it down -- remove noise words and then fetch from its vast database a new set of symbols arranged in its correct syntax (in this case, Javascript used in the threejs library) But the output would have errors or may be incomplete.

Now here comes the interesting part: You ...yes, you, the human who asked the question... you are lot more intelligent. You realized that the code ChatGPT had errors or was partial. For example, when I tried it, the output it gave actually had not added some "import" statements, etc. So you point that out to ChatGPT. So you express yourself back to ChatGPT ("It does not have import statements") and ChatGPT now goes into self-correction mode and adds to its earlier set of instructions kept on the table inside the room!

In short, each user who uses ChatGPT is further training it to become better and better at doing its work. In short, the users of ChatGPT are also an intricate part of the system of ChatGPT. Without us, ChatGPT's capabilities would be limited. It is quite clever.

But is it "sentient"? i.e. does it have true human-like understanding? No.

https://www.youtube.com/watch?v=tBE06SdgzwM

In architecture theory, Christopher Alexander and his colleagues had identified some 253 "syntactic" spatial arrangements that they saw in architecture designs -- and came up with a set of "instructions" on how to "assemble" permutations and combinations of those arrangements to what they thought are quite livable architectural designs. Crudely, it is like assembling a jig-saw puzzle together as per the rules on manipulating those 253 different spatial arrangements. I would not be surprised if ChatGPT uses Christopher Alexander et al's work "A Pattern Language" and uses that and other such research to assemble together architecture designs.

But there is one glaring missing element: Alexander's work did not talk about strict syntactic and linguistic elements. I do agree with Searle that getting the "syntax" right would not get the "semantics" (meaning) also right. But what is also trues is that if you had a loose syntax, it is next to impossible to get any hope of any viable answer right.

Which is where TAD comes in. TAD allows an architect to represent a complex design using a syntactic structure made on top of linguistic ingredients of architecture. To put it simply, I have been pushing for a proposed set of neutral "alphabets" that can be put together as per a grammar to represent any built-form

Nevertheless, Searle's original critcisim of hard AI still remains: The system is most definitely not sentient. It has no real "understanding" But I also agree, that doing architecture design this way will possibly result in less harmful designs out there in the real world -- especially because of its nature of performing self-corrections.

Get TAD here. It is free. www.teamtad.com